devs/foxbox
1
0
Fork 0
forked from len0rd/rockbox
Code Pull requests Activity

Use Tremor IMDCT for WMA. Gives a ~20% speedup on ARM and coldfire. All my test samples gave equivilent or better accuracy (>>16 bit), but there may still be problems. Code becomes somewhat less readable since the IMDCT has bits of Vorbis in it, but this is not a serious issue and the old ffmpeg transform remains in SVN if anyone wants it. Also, WMA now passes Vorbis as the fastest transform codec on ARM, although MPC remains the fastest lossy codec overall.

Browse source 
git-svn-id: svn://svn.rockbox.org/rockbox/trunk@18084 a1c6a512-1295-4272-9138-f99709370657
This commit is contained in:
Michael Giacomelli 2008-07-17 17:43:49 +00:00
parent 5356e6205b
commit 1334738071
15 changed files with 2584 additions and 91 deletions
Download patch file Download diff file Expand all files Collapse all files

35
apps/codecs/libwma/wmadeci.c
View file

@ -28,6 +28,7 @@
#include "wmadec.h"
#include "wmafixed.h"
#include "bitstream.h"
#include "mdct2.h"
#define VLCBITS 7 /*7 is the lowest without glitching*/
@ -589,12 +590,16 @@ int wma_decode_init(WMADecodeContext* s, asf_waveformatex_t *wfx)
}
}
mdct_init_global();
/*Not using the ffmpeg IMDCT anymore*/
/* mdct_init_global();
for(i = 0; i < s->nb_block_sizes; ++i)
{
ff_mdct_init(&s->mdct_ctx[i], s->frame_len_bits - i + 1, 1);
}
*/
/*ffmpeg uses malloc to only allocate as many window sizes as needed. However, we're really only interested in the worst case memory usage.
* In the worst case you can have 5 window sizes, 128 doubling up 2048
@ -1207,8 +1212,10 @@ static int wma_decode_block(WMADecodeContext *s)
n=0;
/*
* Previously the IMDCT was run in 17.15 precision to avoid overflow. However rare files could
* overflow here as well, so switch to 17.15 during coefs calculation.
* The calculation of coefs has a shift right by 2 built in. This prepares samples
* for the Tremor IMDCT which uses a slightly different fixed format then the ffmpeg one.
* If the old ffmpeg imdct is used, each shift storing into coefs should be reduced by 1.
* See SVN logs for details.
*/
@ -1223,7 +1230,7 @@ static int wma_decode_block(WMADecodeContext *s)
/* very low freqs : noise */
for(i = 0;i < s->coefs_start; ++i)
{
*coefs++ = fixmul32( (fixmul32(s->noise_table[s->noise_index],(*exponents++))>>4),Fixed32From64(mult1)) >>1;
*coefs++ = fixmul32( (fixmul32(s->noise_table[s->noise_index],(*exponents++))>>4),Fixed32From64(mult1)) >>2;
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
}
@ -1284,7 +1291,7 @@ static int wma_decode_block(WMADecodeContext *s)
{
noise = s->noise_table[s->noise_index];
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
*coefs++ = fixmul32((fixmul32(*exponents,noise)>>4),Fixed32From64(mult1)) >>1;
*coefs++ = fixmul32((fixmul32(*exponents,noise)>>4),Fixed32From64(mult1)) >>2;
++exponents;
}
}
@ -1299,7 +1306,7 @@ static int wma_decode_block(WMADecodeContext *s)
/*don't forget to renormalize the noise*/
temp1 = (((int32_t)*coefs1++)<<16) + (noise>>4);
temp2 = fixmul32(*exponents, mult>>17);
temp2 = fixmul32(*exponents, mult>>18);
*coefs++ = fixmul32(temp1, temp2);
++exponents;
}
@ -1311,8 +1318,8 @@ static int wma_decode_block(WMADecodeContext *s)
mult2 = fixmul32(mult>>16,exponents[-1]) ; /*the work around for 32.32 vars are getting stupid*/
for (i = 0; i < n; ++i)
{
/*renormalize the noise product and then reduce to 17.15 precison*/
*coefs++ = fixmul32(s->noise_table[s->noise_index],mult2) >>5;
/*renormalize the noise product and then reduce to 14.18 precison*/
*coefs++ = fixmul32(s->noise_table[s->noise_index],mult2) >>6;
s->noise_index = (s->noise_index + 1) & (NOISE_TAB_SIZE - 1);
}
@ -1334,7 +1341,7 @@ static int wma_decode_block(WMADecodeContext *s)
for(i = 0;i < n; ++i)
{
atemp = (coefs1[i] * mult3)>>1;
atemp = (coefs1[i] * mult3)>>2; /*ffmpeg imdct needs 15.17, while tremor 14.18*/
*coefs++=fixmul32(atemp,exponents[i<<bsize>>esize]);
}
n = s->block_len - s->coefs_end[bsize];
@ -1379,9 +1386,13 @@ static int wma_decode_block(WMADecodeContext *s)
n = s->block_len;
n4 = s->block_len >>1;
ff_imdct_calc(&s->mdct_ctx[bsize],
output,
(*(s->coefs))[ch]);
/*faster IMDCT from Vorbis*/
mdct_backward( (1 << (12-bsize)), (int*)(*(s->coefs))[ch], (int*)output);
/*slower but more easily understood IMDCT from FFMPEG*/
//ff_imdct_calc(&s->mdct_ctx[bsize],
// output,
// (*(s->coefs))[ch]);
/* add in the frame */